Create amdgpu_bo_handle from an external GEM handle

To Support Checkpoint and Restore (CRIU) https://github.com/RadeonOpenCompute/criu with AMDGPUs, we are developing a new amdgpu plugin. To copy the VRAM buffer objects we so far used the PROCPIDMEM interface which is very slow in performance. To implement SDMA based operations, we need libdrm support in our plugin but libdrm currently doesn't allow creating amdgpu_bo_handle if the buffer object was not allocated using libdrm. Typical pytorch or tensorflow workloads running on AMD ROCm typically create buffer objects using ROCt library so we need this wrapper API.